Hospital CIOs, IT directors, and radiology managers
What a hospital actually has to provide to run AI orchestration
The honest answer to the question every radiology department asks second, and the one that kills most hospital AI projects: do we need to hire data scientists?
2026-08-01 4 min
No. Not for orchestration. The distinction that matters is between building models and routing to them, and it is the difference between a project that needs a machine learning team and one that needs an afternoon of configuration.
Where the data science requirement actually comes from
The pitch that requires it is "train your own models on your own data". That is a real thing some hospitals do and it is genuinely valuable, but it needs a labelling programme, machine learning engineers, validation infrastructure, and a regulatory pathway of your own. Most radiology departments do not have any of that and are not going to acquire it.
Orchestration inverts the requirement. The models are already built, already trained, and already cleared by their manufacturers. Nothing is trained at the hospital.
What is genuinely needed
- A PACS administrator for the integration window, to configure the DICOM route and test it
- A radiologist to sign off clinical governance, once, for the platform rather than per vendor
- A server, or a cloud tenancy if hosted is acceptable in your jurisdiction
- An information security review, again once rather than per vendor
What is not
- Data scientists or machine learning engineers
- A labelling team or annotation programme
- Training data
- Model tuning or retraining
- A GPU cluster for the orchestration layer itself, though a vendor running on premise will have its own requirements
Where the real cost sits
Integration, not licences. Each AI vendor adopted directly is its own project: PACS work, security review, clinical governance, training, sign-off. Four vendors is four of everything, and it is the reason most departments stop at one.
An orchestration layer makes that a one-time cost, after which adding a vendor is configuration. The saving is not a discount on licences. It is the removal of the cost that stops the second, third and fourth vendor from ever being bought.
If a vendor answers "do we need data scientists" with anything other than a flat no, ask them which specific role does what, and for how long.